Protect the learning inside the work

Account for the expertise the old workflow produced

A firm can improve today’s output while weakening the route through which tomorrow’s experts develop. When AI prepares the case, the junior employee may lose the occasion to identify a contradiction, weigh evidence or defend a recommendation. The question for leaders is how that judgement will now be learned. Faster preparation does not answer it.

Preserve the learning function rather than every old task. Some repetition teaches little; other work builds the mental models needed for unfamiliar cases. Identify the judgement, the conditions in which it matters and the evidence that someone can exercise it. A generic requirement for critical thinking is too broad to guide a workflow redesign.

The risk is a gap between current productivity and future capability: experienced staff keep correcting output while fewer colleagues learn to replace them. This is a hypothesis to test, not an established consequence of AI. The same technology may make expert feedback and practice more accessible. The operating design determines which possibility the firm pursues.

Define a Learning Covenant for each redesigned role

mayden’s Learning Covenant assigns development responsibilities alongside delivery responsibilities. The operating leader names the judgement the role must retain; the employee receives relevant practice and feedback; a competent reviewer assesses readiness for greater responsibility. It is a proposed role-design agreement, not a validated score. Its purpose is to put learning in the operating plan.

For an illustrative commercial analyst, the covenant might require the ability to distinguish a source fact from an assumption, test whether a comparison is valid and explain a recommendation's sensitivity to uncertainty. AI can help retrieve evidence and suggest alternatives. The analyst still needs occasions to form a view before seeing the proposed answer and to defend it against a meaningful challenge. The assessment should examine reasoning on a new case rather than recall of a familiar template.

For a service role, the retained judgement may concern when the standard remedy is inappropriate and how to recognize a vulnerable customer or conflicting evidence. Practice can use carefully prepared cases with known resolution criteria. A senior reviewer should explain the decisive feature, not merely replace the answer. The covenant makes the learning obligation visible in the operating plan, including the time and expertise needed to fulfil it.

Use AI to increase the quality of practice

AI assistance can create useful learning opportunities when the design makes thinking visible. A tool can ask an employee to identify missing evidence, compare alternative explanations or predict the consequence of an action before offering feedback. These are proposed design patterns requiring validation. An apparently helpful explanation can still be wrong, so the practice environment needs trusted reference material and a way for a competent reviewer to examine the feedback.

The support-agent field study reports evidence consistent with worker learning alongside productivity gains. ⁴ It provides evidence against assuming that assistance inevitably reduces skills. The management question is which features of a deployment transmit useful expertise and which merely make users dependent on an answer. That distinction should be investigated through retained performance, transfer to unfamiliar cases and the quality of escalation, not by counting completed learning modules.

Concentrate practice where it supports future responsibility and recovery. Repeating every automated task would consume the capacity the tool releases. Instead, select unfamiliar cases, define the assistance boundary and explain the judgement each exercise develops. Keep routine assistance available where unaided repetition offers little learning value.

Fund teaching as well as approval

A senior professional’s approval and teaching serve different purposes. Approval releases a result; teaching develops someone able to reach and challenge a result later. Separate the two in workload and performance expectations. Otherwise the immediate queue can absorb all senior attention while the firm’s next generation of reviewers remains underdeveloped.

Capture concise explanations of consequential decisions in an authorized knowledge record. Include the context, the evidence that changed the answer and the limits of the reasoning. Avoid presenting a past judgement as a universal rule. A useful case library preserves disagreement and changing conditions as well as successful resolution. Employees should be able to see why superficially similar cases led to different outcomes, which is often where professional judgement becomes visible.

Protect time for this work. If senior staff are measured only on immediate delivery volume, teaching will compete with an urgent queue it rarely wins. The OECD's 2026 skills brief emphasizes that training belongs within a broader response to workplace AI. ⁵ mayden's application is to place capability development inside the operating budget and leadership review. A learning commitment without time, feedback or consequences for neglect is unlikely to survive production pressure.

Measure transfer, calibration and recovery

Assisted performance answers whether the employee and system can produce an acceptable result together. Transfer asks whether the employee can apply the relevant judgement to an unfamiliar case. Calibration asks whether the person recognizes the limits of their own and the system's knowledge. Recovery asks whether they can act when the tool is unavailable or wrong. These are distinct assessment questions. Strong performance on a familiar assisted task should not silently answer all of them.

Exhibit 01 provides a developmental record for an unfamiliar analyst case. The domain owner sets the expected reasoning and assistance boundary. The employee forms a view, then explains what permitted assistance changes. A reviewer records diagnosis, uncertainty, response to a misleading answer and recovery, and gives a reasoned correction. Repeat with a different case after practice. Explain the retained evidence and appeal route: one exercise cannot establish employment fitness, and language or accessibility barriers must not be mistaken for weak domain judgement.

Track development at the role and cohort level as well as individual performance. Are people becoming able to resolve more difficult cases? Is the same small group still rescuing every exception? Are promotions supported by demonstrated judgement? Use multiple observations over time, because a single exercise can reflect familiarity, language or test design. The goal is a credible account of capability development, not a new universal score that gives false precision to a complex learning process.

Promotion criteria should evolve with the redesigned role. An employee who produces polished assisted output may be ready for more responsibility, but the assessment should also examine whether they can frame the problem, identify missing evidence and challenge an attractive wrong answer. Preserve examples of the reasoning behind these judgements so that development conversations are concrete. Where an employee lacks capability, distinguish insufficient practice from an unsuitable assessment or inaccessible tool. The response may be a better learning opportunity rather than a negative performance label. This makes the capability account useful to both the organization and the person whose work is changing.

Keep access and opportunity in the redesign

Exposure estimates from the ILO indicate that many occupations contain tasks potentially affected by generative AI; they do not predict individual job losses. ⁶ Workforce planning should therefore examine how opportunities to learn and advance change within roles. If some employees receive better tools, protected practice and senior attention while others receive only higher output targets, the implementation may widen capability differences even when access to the basic software is nominally equal.

Consider language, accessibility and prior experience in both tools and assessment. A person may understand the domain while struggling with an interface or a training exercise written in a second language. Conversely, fluent interaction with the tool may conceal weak domain judgement. Separate these issues before deciding what development is needed. The organization should provide routes to demonstrate competence that resemble the work and do not reward presentation polish as a substitute for understanding.

UNESCO's guidance advocates human agency and pedagogical validation in educational uses of generative AI. ⁷ Applied to workplace development, this supports an explicit review of who benefits from the learning design and whether it actually teaches the intended capability. It does not make an enterprise programme educationally validated. Leaders should test participation, feedback quality and transfer, then adjust the design when a group is systematically excluded or when apparent progress depends on continued answer provision.

Invest in the next generation of experts

Technology makes some skills obsolete; retaining them all would be wasteful. Preserve judgement needed to frame work, assess evidence, recognize failure and resolve current exceptions. Retire practice that no longer serves those responsibilities, and invest in the capabilities the redesigned role now requires.

Future systems may handle more exceptions. Reassess the covenant as reliability and responsibility change, using evidence rather than a forecast of autonomy. Human practice should support a defined need; equally, an unproven future capability should not justify removing a present recovery skill.

Require each major workflow redesign to explain how its next generation of experts will develop. Approve the retained judgement, practice time and feedback capacity alongside the delivery plan. Then review unfamiliar-case performance and assisted output separately. A stronger result today should not conceal an unfunded capability requirement for tomorrow.